Skip to content

Commit 9998d2c

Browse files
committed
feat: readme-zh
1 parent 6b44019 commit 9998d2c

2 files changed

Lines changed: 142 additions & 1 deletion

File tree

README-ZH.md

Lines changed: 126 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,126 @@
1+
[English](README.md) | **中文**
2+
3+
<p align="center">
4+
<img src="assets/logo.png" alt="Grid Sampler logo" width="140" />
5+
</p>
6+
7+
# Grid Sampler [ICML 2026]
8+
9+
#### See What Matters:面向可泛化视觉-语言-动作模型的可微网格采样剪枝
10+
11+
<p align="center">
12+
<a href="https://scholar.google.com/citations?user=WljJ2HUAAAAJ">Yixu Feng</a><sup>1</sup>,
13+
<a href="https://openreview.net/profile?id=~Zinan_Zhao1">Zinan Zhao</a><sup>2</sup>,
14+
<a href="https://scholar.google.com/citations?user=mBHSbeIAAAAJ">Yanxiang Ma</a><sup>1</sup>,
15+
<a href="https://openreview.net/profile?id=~Chenghao_Xia1">Chenghao Xia</a><sup>3</sup>,
16+
<a href="https://scholar.google.com/citations?user=guY3iCsAAAAJ">Chengbin Du</a><sup>3</sup>,
17+
<a href="https://scholar.google.com/citations?user=m4wbcOsAAAAJ">Yunke Wang</a><sup>1</sup>,
18+
<a href="https://scholar.google.com/citations?user=N4F_3eoAAAAJ">Chang Xu</a><sup>1</sup>
19+
</p>
20+
<p align="center">
21+
<sup>1</sup> 悉尼大学(University of Sydney) &nbsp;·&nbsp;
22+
<sup>2</sup> 香港城市大学(City University of Hong Kong) &nbsp;·&nbsp;
23+
<sup>3</sup> StellarEdge Robotics
24+
</p>
25+
26+
## 演示 Demo 🎞
27+
28+
> **为什么在 README 里看不到内嵌播放器?** GitHub 会过滤 HTML `<video>`,因此无法在项目主页直接嵌入播放。请点击下表中的链接,在 GitHub 文件页使用**内置 MP4 播放器**观看(需已将 `demo/` 下视频推送到默认分支,如 `main`)。
29+
30+
**域内(In-domain)**
31+
32+
| Pen | Pick | Stack |
33+
|:---:|:---:|:---:|
34+
| [▶ 观看 `GridS_demo_pen.mp4`](demo/GridS_demo_pen.mp4) | [▶ 观看 `GridS_demo_pick.mp4`](demo/GridS_demo_pick.mp4) | [▶ 观看 `GridS_demo_stack.mp4`](demo/GridS_demo_stack.mp4) |
35+
36+
**分布外(OOD)**
37+
38+
| Pen (OOD) | Pick (OOD) | Stack (OOD) |
39+
|:---:|:---:|:---:|
40+
| [▶ 观看 `GridS_demo_pen_ood.mp4`](demo/GridS_demo_pen_ood.mp4) | [▶ 观看 `GridS_demo_pick_ood.mp4`](demo/GridS_demo_pick_ood.mp4) | [▶ 观看 `GridS_demo_stack_ood.mp4`](demo/GridS_demo_stack_ood.mp4) |
41+
42+
## 动态 News 🆕
43+
44+
- **2026.05.10** 基于 openpi 的代码已公开。💎
45+
- **2026.05.01** 论文 *See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model* 已被 **ICML 2026** 接收,欢迎基于 Grid Sampler 的后续工作。🔥
46+
47+
## 路线图 To-Do List ✅
48+
49+
- ✅ 发布集成 Grid Sampler 的 openpi 代码。
50+
- ⬜ 发布集成 Grid Sampler 的 X-VLA(Robotwin)代码。
51+
- ⬜ 发布集成 Grid Sampler 的 LeRobot 真实数据代码。
52+
53+
## 结果 Results 📊
54+
55+
<details>
56+
<summary><b>LIBERO:</b></summary>
57+
58+
![Result 1](assets/main_result1.png)
59+
60+
</details>
61+
62+
<details>
63+
<summary><b>真实环境(Real-world):</b></summary>
64+
65+
![Result 2](assets/main_result2.png)
66+
67+
</details>
68+
69+
## 1. 核心思想 Core idea 🌑
70+
71+
![Main idea](assets/core_idea.png)
72+
73+
*GridS 视觉 Token 剪枝框架概览。*
74+
75+
**(a) 标准稠密表示:** 输入图像(*H*<sub>R</sub>、*W*<sub>R</sub> 为原始分辨率)经视觉编码器与 ViT 式嵌入(Dosovitskiy et al., 2020)得到稠密视觉 token(*H* × *W* × *C*),保留完整空间细节。
76+
77+
**(b) GridS Token 剪枝模块:** 在显著区域采样稀疏 token(*K* × *C*),包含两个阶段:(1) 全局坐标预测;(2) 带几何注入的网格采样(Grid Sampling with Geometry Injection)。在 *K**H* × *W* 时,为下游 Transformer 提供高效表示。
78+
79+
## 2. 测试 Testing 🌒
80+
81+
### openpi
82+
83+
可运行流程均在 **openpi** 仓库中说明。请先在 [`openpi/README.md`](openpi/README.md) 完成 **Installation** 环境配置,并以该文档为主索引:
84+
85+
- **快速策略自检(无真机):****Running Inference for a Pre-Trained Model** 一节中的 **「Test inference without a robot」** 段落,指向 [`openpi/examples/simple_client/README.md`](openpi/examples/simple_client/README.md)
86+
- **Notebook / 通用推理:** 同上章节及其中提到的 `examples/inference.ipynb`
87+
- **LIBERO 仿真评测:** [`openpi/examples/libero/README.md`](openpi/examples/libero/README.md)(主 README 微调模型表中亦有 LIBERO 相关说明)。
88+
- **ALOHA:** 仿真见 [`openpi/examples/aloha_sim/README.md`](openpi/examples/aloha_sim/README.md);真机见 [`openpi/examples/aloha_real/README.md`](openpi/examples/aloha_real/README.md);汇总亦见 [`openpi/README.md`](openpi/README.md)**More Examples**
89+
90+
## 3. 微调 Finetuning 🌓
91+
92+
### openpi
93+
94+
微调流程遵循 **openpi** 官方路径,完整说明见 [`openpi/README.md`](openpi/README.md) 中的 **[Fine-Tuning Base Models on Your Own Data](openpi/README.md#fine-tuning-base-models-on-your-own-data)**。简要步骤:
95+
96+
1. **数据:** 转为 LeRobot 数据集(LIBERO 示例脚本:[`openpi/examples/libero/convert_libero_data_to_lerobot.py`](openpi/examples/libero/convert_libero_data_to_lerobot.py));若仅使用上游配置中已绑定的 LIBERO 数据,通常可跳过自行转换。
97+
2. **配置:** 数据变换与 `TrainConfig`[`openpi/src/openpi/training/config.py`](openpi/src/openpi/training/config.py)(LIBERO 策略见 [`openpi/src/openpi/policies/libero_policy.py`](openpi/src/openpi/policies/libero_policy.py))。使用 **Grid Sampler** 时请选择模型侧 `grid=True` 的训练配置(如 `pi0_libero_grid*``pi05_libero_grid*` 等)。
98+
3. **训练(JAX):** 先按文档计算归一化统计量,例如 `uv run scripts/compute_norm_stats.py --config-name <your_config>`,再 `uv run scripts/train.py <your_config> --exp-name=...`(GPU 显存相关见同文档中的 `XLA_PYTHON_CLIENT_MEM_FRACTION` 等说明)。
99+
4. **PyTorch 路线:** 若使用 PyTorch 实现,见同 README 的 **[PyTorch Support](openpi/README.md#pytorch-support)**(环境、`train_pytorch.py`、JAX→PyTorch 权重转换等)。
100+
101+
部署微调后的 checkpoint:见上述章节中的 **Spinning up a policy server and running inference**,以及 [`openpi/scripts/serve_policy.py`](openpi/scripts/serve_policy.py) 与 LIBERO 客户端文档 [`openpi/examples/libero/README.md`](openpi/examples/libero/README.md)
102+
103+
## 4. 联系方式 Contacts 🌔
104+
105+
如有问题,欢迎发邮件或在仓库中提 Issue,感谢反馈与贡献。
106+
107+
**Yixu Feng:** [yfen0429@sydney.edu.au](yfen0429@sydney.edu.au)[fedioryf@gmail.com](fedioryf@gmail.com)
108+
109+
**Zinan Zhao:** [zhao48zinan@gmail.com](zhao48zinan@gmail.com)
110+
111+
也可扫描下方微信二维码联系:
112+
113+
<img src="assets/wechat.jpg" alt="微信二维码" width="220">
114+
115+
## 5. 引用 Citation 🌕
116+
117+
若本工作对您的研究有帮助,欢迎引用:
118+
119+
```
120+
@inproceedings{feng2026gridsampler,
121+
title = {See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model},
122+
author = {Feng, Yixu and Zhao, Zinan and Ma, Yanxiang and Xia, Chenghao and Du, Chengbin and Wang, Yunke and Xu, Chang},
123+
booktitle = {Forty-Third International Conference on Machine Learning (ICML)},
124+
year = {2026}
125+
}
126+
```

README.md

Lines changed: 16 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4,9 +4,24 @@
44

55
# Grid Sampler [ICML 2026]
66

7+
[简体中文](README-ZH.md)
8+
79
#### See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model
810

9-
[Yixu Feng](https://scholar.google.com/citations?user=WljJ2HUAAAAJ), [Zinan Zhao](https://openreview.net/profile?id=~Zinan_Zhao1), [Yanxiang Ma](https://scholar.google.com/citations?user=mBHSbeIAAAAJ), [Chenghao Xia](https://openreview.net/profile?id=~Chenghao_Xia1), [Chengbin Du](https://scholar.google.com/citations?user=guY3iCsAAAAJ), [Yunke Wang](https://scholar.google.com/citations?user=m4wbcOsAAAAJ), [Chang Xu](https://scholar.google.com/citations?user=N4F_3eoAAAAJ)
11+
<p align="center">
12+
<a href="https://scholar.google.com/citations?user=WljJ2HUAAAAJ">Yixu Feng</a><sup>1</sup>,
13+
<a href="https://openreview.net/profile?id=~Zinan_Zhao1">Zinan Zhao</a><sup>2</sup>,
14+
<a href="https://scholar.google.com/citations?user=mBHSbeIAAAAJ">Yanxiang Ma</a><sup>1</sup>,
15+
<a href="https://openreview.net/profile?id=~Chenghao_Xia1">Chenghao Xia</a><sup>3</sup>,
16+
<a href="https://scholar.google.com/citations?user=guY3iCsAAAAJ">Chengbin Du</a><sup>3</sup>,
17+
<a href="https://scholar.google.com/citations?user=m4wbcOsAAAAJ">Yunke Wang</a><sup>1</sup>,
18+
<a href="https://scholar.google.com/citations?user=N4F_3eoAAAAJ">Chang Xu</a><sup>1</sup>
19+
</p>
20+
<p align="center">
21+
<sup>1</sup> University of Sydney &nbsp;·&nbsp;
22+
<sup>2</sup> City University of Hong Kong &nbsp;·&nbsp;
23+
<sup>3</sup> StellarEdge Robotics
24+
</p>
1025

1126
## Demo 🎞
1227

0 commit comments

Comments
 (0)